Step 1: Dataset Exploration¶
In [1]:
%matplotlib inline
import numpy as np
import matplotlib
matplotlib.use('TkAgg', warn = False)
from matplotlib import pyplot
import tensorflow as tf
In [2]:
import pickle
def load_pickled_data(file, columns):
"""
Loads pickled training and test data.
Parameters
----------
file :
Name of the pickle file.
columns : list of strings, optional, defaults to `['features', 'labels']`
List of columns in pickled data we're interested in.
Returns
-------
A tuple of datasets for given columns.
"""
with open(file, mode='rb') as f:
dataset = pickle.load(f)
return tuple(map(lambda c: dataset[c], columns))
In [3]:
from pandas.io.parsers import read_csv
signnames = read_csv("signnames.csv").values[:, 1]
train_dataset_file = "traffic-signs-data/train.p"
test_dataset_file = "traffic-signs-data/test.p"
train_balanced_dataset_file = "traffic-signs-data/train_balanced.p"
X_train, y_train, sizes, positions = load_pickled_data(train_dataset_file, ['features', 'labels', 'sizes', 'coords'])
X_test, y_test = load_pickled_data(test_dataset_file, ['features', 'labels'])
n_train = y_train.shape[0]
n_test = y_test.shape[0]
image_shape = X_train[0].shape
image_size = image_shape[0]
sign_classes, class_indices, class_counts = np.unique(y_train, return_index = True, return_counts = True)
n_classes = class_counts.shape[0]
print("Number of training examples =", n_train)
print("Number of testing examples =", n_test)
print("Image data shape =", image_shape)
print("Number of classes =", n_classes)
Original photos¶
Let's first explore data about original photos. Who knows — maybe we can spot some data correlations with signs positions in the photos or positions across specific sign classes.
Let's first plot histograms of original photos' sizes.
In [4]:
width_fraction = (positions[:, 2] - positions[:, 0]) / sizes[:, 0]
height_fraction = (positions[:, 3] - positions[:, 1]) / sizes[:, 1]
sizes_figure = pyplot.figure(figsize = (10, 4))
axis = sizes_figure.add_subplot(1, 2, 1)
axis.hist(sizes[:, 0])
pyplot.title('Original photos\'s widths')
pyplot.xlabel('Width')
pyplot.ylabel('Photos')
axis = sizes_figure.add_subplot(1, 2, 2)
axis.hist(sizes[:, 1])
pyplot.title('Original photos\'s heights')
pyplot.xlabel('Height')
pyplot.ylabel('Photos')
pyplot.show()
Doesn't look too promising — apparently photos are of all different sizes, majority ranging from ~30 to ~150 px both width and height. Actually both histograms looks surprisingly similar, I bet aspect ratios of all photos are close to 1. Let's check it out!
In [5]:
sizes_figure = pyplot.figure()
axis = sizes_figure.add_subplot(1, 1, 1)
axis.hist(sizes[:, 0]/sizes[:, 1])
pyplot.title('Photos\' aspect ratios distribution')
pyplot.xlabel('Aspect ratio')
pyplot.ylabel('Photos')
pyplot.show()
There you go! Vast majority of original photos had aspect ratio ranging from 0.9 to 1.2, e.g. almost all photos were more or less square.
Let's check how much area of the original image did signs occupy? In other words, to which extent did we have to crop the original photo to get the sign out of it. Again, let's check for both width and height by calculating sign dimensions as fractions of original photo dimensions.
In [6]:
width_fraction = (positions[:, 2] - positions[:, 0]) / sizes[:, 0]
height_fraction = (positions[:, 3] - positions[:, 1]) / sizes[:, 1]
sizes_figure = pyplot.figure(figsize = (10, 4))
axis = sizes_figure.add_subplot(1, 2, 1)
axis.hist(width_fraction)
pyplot.title('Sign width / original photo width')
pyplot.xlabel('Width fraction')
pyplot.ylabel('Photos')
axis = sizes_figure.add_subplot(1, 2, 2)
axis.hist(height_fraction)
pyplot.title('Sign height / original photo height')
pyplot.xlabel('Height fraction')
pyplot.ylabel('Photos')
pyplot.show()
In [7]:
from scipy import stats
def print_stats(array, label):
_, minmax, mean, variance, _, _ = stats.describe(array)
margin = int((40 - len(label)) / 2)
for i in range(margin * 2):
if i == margin: print(label.upper(), end="")
print("=", end="")
print()
print(" %-*s %s" % (15, "Min:", str(minmax[0])))
print(" %-*s %s" % (15, "Max:", str(minmax[1])))
print(" %-*s %s" % (15, "Mean:", str(mean)))
print(" %-*s %s" % (15, "Variance:", str(variance)))
print()
print_stats(width_fraction, " sign_width / photo_width ")
print_stats(height_fraction, " sign_height / photo_height ")
Not really promising either. All photos had almost nothing but signs in them, for both width and height signs occupied at least half of the dimension, and for majority of samples signs occupied around 75% of the photo.
Now, if we found any uneven distributions in the dataset I would be tempted to try and analyze it on per-class basis trying to find any patterns like specific signs appearing in some parts of the image (in case all data was collected from a similar relative point). Although for now data looks pretty homogeneous and I doubt there will be any meaningful patterns related to specific classes based on the original samples' images.
Sample images¶
Ok, let's check out the sign samples. We will go through all dataset classes, noting number of samples and plotting 10 random images representing each class.
In [8]:
import random
col_width = max(len(name) for name in signnames)
for c, c_index, c_count in zip(sign_classes, class_indices, class_counts):
print("Class %i: %-*s %s samples" % (c, col_width, signnames[c], str(c_count)))
fig = pyplot.figure(figsize = (6, 1))
fig.subplots_adjust(left = 0, right = 1, bottom = 0, top = 1, hspace = 0.05, wspace = 0.05)
random_indices = random.sample(range(c_index, c_index + c_count), 10)
for i in range(10):
axis = fig.add_subplot(1, 10, i + 1, xticks=[], yticks=[])
axis.imshow(X_train[random_indices[i]])
pyplot.show()
print("--------------------------------------------------------------------------------------\n")
Some classes are highly underrepresented. Overall amount of data is disappointing: some classes have as little as 210 samples, this won't be enough for most of the models to generalise well.
Images with some signs are horizontally symmetrical (like Bumpy road or Ahead only), and can be simply flipped horizontally, thus allowing us to get twice as much data for these classes.
Other signs come in kind of interchageable pairs, like Keep right and Keep left: those signs can be flipped and assigned to a paired class. In some cases (like Keep right and Keep left) we increase the number of samples for Keep left from 300 to 2370!
CNNs have built-in invariance to small translations, scaling and rotations. The training set doesn't seem to contain those deformations, so we will add those in our data augmentation step as well.
Step 2: Design and Test a Model Architecture¶
Some useful logging routines.
In [4]:
import sys
import time
def get_time_hhmmss(start = None):
"""
Calculates time since `start` and formats as a string.
"""
if start is None:
return time.strftime("%Y/%m/%d %H:%M:%S")
end = time.time()
m, s = divmod(end - start, 60)
h, m = divmod(m, 60)
time_str = "%02d:%02d:%02d" % (h, m, s)
return time_str
# Print iterations progress
def print_progress(iteration, total):
"""
Call in a loop to create terminal progress bar
@params:
iteration - Required : current iteration (Int)
total - Required : total iterations (Int)
"""
str_format = "{0:.0f}"
percents = str_format.format(100 * (iteration / float(total)))
filled_length = int(round(100 * iteration / float(total)))
bar = 'â–ˆ' * filled_length + '-' * (100 - filled_length)
sys.stdout.write('\r |%s| %s%%' % (bar, percents)),
if iteration == total:
sys.stdout.write('\n')
sys.stdout.flush()
Preprocessing routines.
In [28]:
from sklearn.utils import shuffle
from skimage import exposure
import warnings
num_classes = 43
def preprocess_dataset(X, y = None):
"""
Performs feature scaling, one-hot encoding of labels and shuffles the data if labels are provided.
Assumes original dataset is sorted by labels.
Parameters
----------
X : ndarray
Dataset array containing feature examples.
y : ndarray, optional, defaults to `None`
Dataset labels in index form.
Returns
-------
A tuple of X and y.
"""
print("Preprocessing dataset with {} examples:".format(X.shape[0]))
#Convert to grayscale, e.g. single channel Y
X = 0.299 * X[:, :, :, 0] + 0.587 * X[:, :, :, 1] + 0.114 * X[:, :, :, 2]
#Scale features to be in [0, 1]
X = (X / 255.).astype(np.float32)
for i in range(X.shape[0]):
with warnings.catch_warnings():
warnings.simplefilter("ignore")
X[i] = exposure.equalize_adapthist(X[i])
print_progress(i + 1, X.shape[0])
if y is not None:
# Convert to one-hot encoding. Convert back with `y = y.nonzero()[1]`
y = np.eye(num_classes)[y]
X, y = shuffle(X, y)
# Add a single grayscale channel
X = X.reshape(X.shape + (1,))
return X, y
def class_name(one_hot):
return signnames[one_hot.nonzero()[0][0]]
def load_and_process_data(pickled_data_file):
"""
Loads pickled data and preprocesses images and labels by scaling features,
shuffling the data and applying one-hot encoding to labels.
Parameters
----------
pickled_data_file : string
Pickled data file name.
Returns
-------
A tuple of X and y containing preloaded and preprocessed features and labels respectively.
"""
X, y = load_pickled_data(pickled_data_file, columns = ['features', 'labels'])
X, y = preprocess_dataset(X, y)
return (X, y)
Data augmentation.
In [ ]:
from nolearn.lasagne import BatchIterator
from skimage.transform import rotate
from skimage.transform import warp
from skimage.transform import ProjectiveTransform
import random
class AugmentedSignsBatchIterator(BatchIterator):
"""
Iterates over dataset in batches.
Allows images augmentation by randomly rotating, applying projection,
adjusting gamma, blurring, adding noize and flipping horizontally.
"""
def __init__(self, batch_size, shuffle = False, seed = 42, p = 0.5, intensity = 0.5):
"""
Initialises an instance with usual iterating settings, as well as data augmentation coverage
and augmentation intensity.
Parameters
----------
batch_size:
Size of the iteration batch.
shuffle :
Flag indicating if we need to shuffle the data.
seed :
Random seed.
p :
Probability of augmenting a single example, should be in a range of [0, 1] .
Defines data augmentation coverage.
intensity :
Augmentation intensity, should be in a [0, 1] range.
Returns
-------
New batch iterator instance.
"""
super(AugmentedSignsBatchIterator, self).__init__(batch_size, shuffle, seed)
self.p = p
self.intensity = intensity
def transform(self, Xb, yb):
"""
Applies a pipeline of randomised transformations for data augmentation.
"""
Xb, yb = super(AugmentedSignsBatchIterator, self).transform(
Xb if yb is None else Xb.copy(),
yb
)
if yb is not None:
batch_size = Xb.shape[0]
image_size = Xb.shape[1]
Xb = self.rotate(Xb, batch_size)
Xb = self.apply_projection_transform(Xb, batch_size, image_size)
return Xb, yb
def rotate(self, Xb, batch_size):
"""
Applies random rotation in a defined degrees range to a random subset of images.
Range itself is subject to scaling depending on augmentation intensity.
"""
for i in np.random.choice(batch_size, int(batch_size * self.p), replace = False):
delta = 30. * self.intensity # scale by self.intensity
Xb[i] = rotate(Xb[i], random.uniform(-delta, delta), mode = 'edge')
return Xb
def apply_projection_transform(self, Xb, batch_size, image_size):
"""
Applies projection transform to a random subset of images. Projection margins are randomised in a range
depending on the size of the image. Range itself is subject to scaling depending on augmentation intensity.
"""
d = image_size * 0.3 * self.intensity
for i in np.random.choice(batch_size, int(batch_size * self.p), replace = False):
tl_top = random.uniform(-d, d) # Top left corner, top margin
tl_left = random.uniform(-d, d) # Top left corner, left margin
bl_bottom = random.uniform(-d, d) # Bottom left corner, bottom margin
bl_left = random.uniform(-d, d) # Bottom left corner, left margin
tr_top = random.uniform(-d, d) # Top right corner, top margin
tr_right = random.uniform(-d, d) # Top right corner, right margin
br_bottom = random.uniform(-d, d) # Bottom right corner, bottom margin
br_right = random.uniform(-d, d) # Bottom right corner, right margin
transform = ProjectiveTransform()
transform.estimate(np.array((
(tl_left, tl_top),
(bl_left, image_size - bl_bottom),
(image_size - br_right, image_size - br_bottom),
(image_size - tr_right, tr_top)
)), np.array((
(0, 0),
(0, image_size),
(image_size, image_size),
(image_size, 0)
)))
Xb[i] = warp(Xb[i], transform, output_shape=(image_size, image_size), order = 1, mode = 'edge')
return Xb
Let's test data augmentation by plotting first 5 examples next to the original images (original image is on the left).
In [22]:
X_train, y_train = load_pickled_data("traffic-signs-data/train.p", columns = ['features', 'labels'])
X_train = X_train / 255.
batch_iterator = AugmentedSignsBatchIterator(batch_size = 5, p = 1.0, intensity = 0.75)
for x_batch, y_batch in batch_iterator(X_train, y_train):
for i in range(5):
# plot two images:
fig = pyplot.figure(figsize=(3, 1))
axis = fig.add_subplot(1, 2, 1, xticks=[], yticks=[])
axis.imshow(X_train[i])
axis = fig.add_subplot(1, 2, 2, xticks=[], yticks=[])
axis.imshow(x_batch[i])
pyplot.show()
break
A couple of functions for preprocessing our dataset. We may only need to call them once, then pickle resulting balanced dataset and keep using that pickle from then on.
In [7]:
def flip_extend(X, y):
"""
Extends existing images dataset by flipping images of some classes. As some images would still belong
to same class after flipping we extend such classes with flipped images. Images of other would toggle
between two classes when flipped, so for those we extend existing datasets as well.
Parameters
----------
X : ndarray
Dataset array containing feature examples.
y : ndarray, optional, defaults to `None`
Dataset labels in index form.
Returns
-------
A tuple of X and y.
"""
# Classes of signs that, when flipped horizontally, should still be classified as the same class
self_flippable = np.array([11, 12, 13, 15, 17, 18, 22, 26, 30, 35])
# Classes of signs that, when flipped vertically, should still be classified as the same class
self_flippable_vertically = np.array([1, 5, 12, 15, 17])
# Classes of signs that, when flipped horizontally and then vertically, should still be classified as the same class
self_flippable_both = np.array([32, 40])
# Classes of signs that, when flipped horizontally, would still be meaningful, but should be classified as some other class
cross_flippable = np.array([
[19, 20],
[33, 34],
[36, 37],
[38, 39],
[20, 19],
[34, 33],
[37, 36],
[39, 38],
])
num_classes = 43
X_extended = np.empty([0, X.shape[1], X.shape[2], X.shape[3]], dtype = X.dtype)
y_extended = np.empty([0], dtype = y.dtype)
for c in range(num_classes):
# First copy existing data for this class
X_extended = np.append(X_extended, X[y == c], axis = 0)
# If we can flip images of this class horizontally and they would still belong to said class...
if c in self_flippable:
# ...Copy their flipped versions into extended array.
X_extended = np.append(X_extended, X[y == c][:, :, ::-1, :], axis = 0)
# If we can flip images of this class horizontally and they would belong to other class...
if c in cross_flippable[:, 0]:
# ...Copy flipped images of that other class to the extended array.
flip_class = cross_flippable[cross_flippable[:, 0] == c][0][1]
X_extended = np.append(X_extended, X[y == flip_class][:, :, ::-1, :], axis = 0)
# If we can flip images of this class horizontally and they would still belong to said class...
if c in self_flippable_vertically:
# ...Copy their flipped versions into extended array.
X_extended = np.append(X_extended, X_extended[y == c][:, ::-1, :, :], axis = 0)
# If we can flip images of this class horizontally and they would still belong to said class...
if c in self_flippable_both:
# ...Copy their flipped versions into extended array.
X_extended = np.append(X_extended, X_extended[y == c][:, ::-1, ::-1, :], axis = 0)
# Fill labels for added images set to current class.
y_extended = np.append(y_extended, np.full((X_extended.shape[0] - y_extended.shape[0]), c, dtype = int))
return (X_extended, y_extended)
def extend_balancing_classes(X, y, aug_intensity = 0.5, counts = None):
"""
Extends dataset by duplicating existing images while applying data augmentation pipeline.
Number of generated examples for each class may be provided in `counts`.
Parameters
----------
X : ndarray
Dataset array containing feature examples.
y : ndarray, optional, defaults to `None`
Dataset labels in index form.
aug_intensity :
Intensity of augmentation, must be in [0, 1] range.
counts :
Number of elements for each class.
Returns
-------
A tuple of X and y.
"""
num_classes = 43
_, class_counts = np.unique(y, return_counts = True)
max_c = max(class_counts)
total = max_c * num_classes if counts is None else np.sum(counts)
X_extended = np.empty([0, X.shape[1], X.shape[2], X.shape[3]], dtype = np.float32)
y_extended = np.empty([0], dtype = y.dtype)
print("Extending dataset using augmented data (intensity = {}):".format(aug_intensity))
for c, c_count in zip(range(num_classes), class_counts):
# How many examples should there be eventually for this class:
max_c = max_c if counts is None else counts[c]
# First copy existing data for this class
X_source = (X[y == c] / 255.).astype(np.float32)
y_source = y[y == c]
X_extended = np.append(X_extended, X_source, axis = 0)
for i in range((max_c // c_count) - 1):
batch_iterator = AugmentedSignsBatchIterator(batch_size = X_source.shape[0], p = 1.0, intensity = aug_intensity)
for x_batch, _ in batch_iterator(X_source, y_source):
X_extended = np.append(X_extended, x_batch, axis = 0)
print_progress(X_extended.shape[0], total)
batch_iterator = AugmentedSignsBatchIterator(batch_size = max_c % c_count, p = 1.0, intensity = aug_intensity)
for x_batch, _ in batch_iterator(X_source, y_source):
X_extended = np.append(X_extended, x_batch, axis = 0)
print_progress(X_extended.shape[0], total)
break
# Fill labels for added images set to current class.
added = X_extended.shape[0] - y_extended.shape[0]
y_extended = np.append(y_extended, np.full((added), c, dtype = int))
return ((X_extended * 255.).astype(np.uint8), y_extended)
Struct to organise some of the model parameters.
In [8]:
from collections import namedtuple
Parameters = namedtuple('Parameters', [
# Data parameters
'num_classes', 'image_size',
# Training parameters
'batch_size', 'max_epochs', 'log_epoch', 'print_epoch',
# Optimisations
'learning_rate_decay', 'learning_rate',
'l2_reg_enabled', 'l2_lambda',
'early_stopping_enabled', 'early_stopping_patience',
'resume_training',
# Layers architecture
'conv1_k', 'conv1_d', 'conv1_p',
'conv2_k', 'conv2_d', 'conv2_p',
'conv3_k', 'conv3_d', 'conv3_p',
'fc4_size', 'fc4_p'
])
General routine for managing model paths, scopes and identifiers.
In [9]:
import os
class Paths(object):
"""
Provides easy access to common paths we use for persisting
the data associated with model training.
"""
def __init__(self, params):
"""
Initialises a new `Paths` instance and creates corresponding folders if needed.
Parameters
----------
params : Parameters
Structure (`namedtuple`) containing model parameters.
"""
self.model_name = self.get_model_name(params)
self.var_scope = self.get_variables_scope(params)
self.root_path = os.getcwd() + "/models/" + self.model_name + "/"
self.model_path = self.get_model_path()
self.train_history_path = self.get_train_history_path()
self.learning_curves_path = self.get_learning_curves_path()
os.makedirs(self.root_path, exist_ok = True)
def get_model_name(self, params):
"""
Generates a model name with some of the crucial model parameters encoded into the name.
Parameters
----------
params : Parameters
Structure (`namedtuple`) containing model parameters.
Returns
-------
Model name.
"""
# We will encode model settings in its name: architecture, optimisations applied, etc.
model_name = "k{}d{}p{}_k{}d{}p{}_k{}d{}p{}_fc{}p{}".format(
params.conv1_k, params.conv1_d, params.conv1_p,
params.conv2_k, params.conv2_d, params.conv2_p,
params.conv3_k, params.conv3_d, params.conv3_p,
params.fc4_size, params.fc4_p
)
model_name += "_lrdec" if params.learning_rate_decay else "_no-lrdec"
model_name += "_l2" if params.l2_reg_enabled else "_no-l2"
return model_name
def get_variables_scope(self, params):
"""
Generates a model variable scope with some of the crucial model parameters encoded.
Parameters
----------
params : Parameters
Structure (`namedtuple`) containing model parameters.
Returns
-------
Variables scope name.
"""
# We will encode model settings in its name: architecture, optimisations applied, etc.
var_scope = "k{}d{}_k{}d{}_k{}d{}_fc{}_fc0".format(
params.conv1_k, params.conv1_d,
params.conv2_k, params.conv2_d,
params.conv3_k, params.conv3_d,
params.fc4_size
)
return var_scope
def get_model_path(self):
"""
Generates path to the model file.
Returns
-------
Model file path.
"""
return self.root_path + "model.ckpt"
def get_train_history_path(self):
"""
Generates path to the train history file.
Returns
-------
Train history file path.
"""
return self.root_path + "train_history"
def get_learning_curves_path(self):
"""
Generates path to the learning curves graph file.
Returns
-------
Learning curves file path.
"""
return self.root_path + "learning_curves.png"
Early stopping.
In [10]:
class EarlyStopping(object):
"""
Provides early stopping functionality. Keeps track of model accuracy,
and if it doesn't improve over time restores last best performing
parameters.
"""
def __init__(self, saver, session, patience = 100, minimize = True):
"""
Initialises a `EarlyStopping` isntance.
Parameters
----------
saver :
TensorFlow Saver object to be used for saving and restoring model.
session :
TensorFlow Session object containing graph where model is restored.
patience :
Early stopping patience. This is the number of epochs we wait for
accuracy to start improving again before stopping and restoring
previous best performing parameters.
Returns
-------
New instance.
"""
self.minimize = minimize
self.patience = patience
self.saver = saver
self.session = session
self.best_monitored_value = np.inf if minimize else 0.
self.best_monitored_epoch = 0
self.restore_path = None
def __call__(self, value, epoch):
"""
Checks if we need to stop and restores the last well performing values if we do.
Parameters
----------
value :
Last epoch monitored value.
epoch :
Last epoch number.
Returns
-------
`True` if we waited enough and it's time to stop and we restored the
best performing weights, or `False` otherwise.
"""
if (self.minimize and value < self.best_monitored_value) or (not self.minimize and value > self.best_monitored_value):
self.best_monitored_value = value
self.best_monitored_epoch = epoch
self.restore_path = self.saver.save(self.session, os.getcwd() + "/early_stopping_checkpoint")
elif self.best_monitored_epoch + self.patience < epoch:
if self.restore_path != None:
self.saver.restore(self.session, self.restore_path)
else:
print("ERROR: Failed to restore session")
return True
return False
Logger to keep track of the training even if no browser is connected to kernel.
In [11]:
from cloudlog import CloudLog
class ModelCloudLog(CloudLog):
def log_parameters(self, params, train_size, valid_size, test_size):
"""
Logs model parameters to console and appends the same text representation to the log file.
Parameters
----------
params : Parameters
Structure (`namedtuple`) containing model parameters.
train_size: int
Size of the training dataset.
valid_size: int
Size of the training dataset.
test_size : int
Size of the training dataset.
"""
if params.resume_training:
self("=============================================")
self("============= RESUMING TRAINING =============")
self("=============================================")
self("=================== DATA ====================")
self(" Training set: {} examples".format(train_size))
self(" Validation set: {} examples".format(valid_size))
self(" Testing set: {} examples".format(test_size))
self(" Batch size: {}".format(params.batch_size))
self("=================== MODEL ===================")
self("--------------- ARCHITECTURE ----------------")
self(" %-*s %-*s %-*s %-*s" % (10, "", 10, "Type", 8, "Size", 15, "Dropout (keep p)"))
self(" %-*s %-*s %-*s %-*s" % (10, "Layer 1", 10, "{}x{} Conv".format(params.conv1_k, params.conv1_k), 8, str(params.conv1_d), 15, str(params.conv1_p)))
self(" %-*s %-*s %-*s %-*s" % (10, "Layer 2", 10, "{}x{} Conv".format(params.conv2_k, params.conv2_k), 8, str(params.conv2_d), 15, str(params.conv2_p)))
self(" %-*s %-*s %-*s %-*s" % (10, "Layer 3", 10, "{}x{} Conv".format(params.conv3_k, params.conv3_k), 8, str(params.conv3_d), 15, str(params.conv3_p)))
self(" %-*s %-*s %-*s %-*s" % (10, "Layer 4", 10, "FC", 8, str(params.fc4_size), 15, str(params.fc4_p)))
self("---------------- PARAMETERS -----------------")
self(" Learning rate decay: " + ("Enabled" if params.learning_rate_decay else "Disabled (rate = {})".format(params.learning_rate)))
self(" L2 Regularization: " + ("Enabled (lambda = {})".format(params.l2_lambda) if params.l2_reg_enabled else "Disabled"))
self(" Early stopping: " + ("Enabled (patience = {})".format(params.early_stopping_patience) if params.early_stopping_enabled else "Disabled"))
self(" Keep training old model: " + ("Enabled" if params.resume_training else "Disabled"))
Model architecture
In [12]:
def fully_connected(input, size):
"""
Performs a single fully connected layer pass, e.g. returns `input * weights + bias`.
"""
weights = tf.get_variable( 'weights',
shape = [input.get_shape()[1], size],
initializer = tf.contrib.layers.xavier_initializer()
)
biases = tf.get_variable( 'biases',
shape = [size],
initializer = tf.constant_initializer(0.0)
)
return tf.matmul(input, weights) + biases
def fully_connected_relu(input, size):
return tf.nn.relu(fully_connected(input, size))
def conv_relu(input, kernel_size, depth):
"""
Performs a single convolution layer pass.
"""
weights = tf.get_variable( 'weights',
shape = [kernel_size, kernel_size, input.get_shape()[3], depth],
initializer = tf.contrib.layers.xavier_initializer()
)
biases = tf.get_variable( 'biases',
shape = [depth],
initializer = tf.constant_initializer(0.0)
)
conv = tf.nn.conv2d(input, weights,
strides = [1, 1, 1, 1], padding = 'SAME')
return tf.nn.relu(conv + biases)
def pool(input, size):
"""
Performs a max pooling layer pass.
"""
return tf.nn.max_pool(
input,
ksize = [1, size, size, 1],
strides = [1, size, size, 1],
padding = 'SAME'
)
def model_pass(input, params, is_training):
"""
Performs a full model pass.
Parameters
----------
input : Tensor
NumPy array containing a batch of examples.
params : Parameters
Structure (`namedtuple`) containing model parameters.
is_training : Tensor of type tf.bool
Flag indicating if we are training or not (e.g. whether to use dropout).
Returns
-------
Tensor with predicted logits.
"""
# Convolutions
with tf.variable_scope('conv1'):
conv1 = conv_relu(input, kernel_size = params.conv1_k, depth = params.conv1_d)
with tf.variable_scope('pool1'):
pool1 = pool(conv1, size = 2)
pool1 = tf.cond(is_training, lambda: tf.nn.dropout(pool1, keep_prob = params.conv1_p), lambda: pool1)
with tf.variable_scope('conv2'):
conv2 = conv_relu(pool1, kernel_size = params.conv2_k, depth = params.conv2_d)
with tf.variable_scope('pool2'):
pool2 = pool(conv2, size = 2)
pool2 = tf.cond(is_training, lambda: tf.nn.dropout(pool2, keep_prob = params.conv2_p), lambda: pool2)
with tf.variable_scope('conv3'):
conv3 = conv_relu(pool2, kernel_size = params.conv3_k, depth = params.conv3_d)
with tf.variable_scope('pool3'):
pool3 = pool(conv3, size = 2)
pool3 = tf.cond(is_training, lambda: tf.nn.dropout(pool3, keep_prob = params.conv3_p), lambda: pool3)
# Fully connected
# 1st stage output
pool1 = pool(pool1, size = 4)
shape = pool1.get_shape().as_list()
pool1 = tf.reshape(pool1, [-1, shape[1] * shape[2] * shape[3]])
# 2nd stage output
pool2 = pool(pool2, size = 2)
shape = pool2.get_shape().as_list()
pool2 = tf.reshape(pool2, [-1, shape[1] * shape[2] * shape[3]])
# 3rd stage output
shape = pool3.get_shape().as_list()
pool3 = tf.reshape(pool3, [-1, shape[1] * shape[2] * shape[3]])
flattened = tf.concat(1, [pool1, pool2, pool3])
with tf.variable_scope('fc4'):
fc4 = fully_connected_relu(flattened, size = params.fc4_size)
fc4 = tf.cond(is_training, lambda: tf.nn.dropout(fc4, keep_prob = params.fc4_p), lambda: fc4)
with tf.variable_scope('out'):
logits = fully_connected(fc4, size = params.num_classes)
return logits
Plotting learning curves.
In [13]:
def plot_curve(axis, params, train_column, valid_column, linewidth = 2, train_linestyle = "b-", valid_linestyle = "g-"):
model_history = np.load(Paths(params).train_history_path + ".npz")
train_values = model_history[train_column]
valid_values = model_history[valid_column]
epochs = train_values.shape[0]
x_axis = np.arange(epochs)
axis.plot(x_axis[train_values > 0], train_values[train_values > 0], train_linestyle, linewidth=linewidth, label="train")
axis.plot(x_axis[valid_values > 0], valid_values[valid_values > 0], valid_linestyle, linewidth=linewidth, label="valid")
return epochs
# Plots history of learning curves for a specific model.
def plot_learning_curves(params):
curves_figure = pyplot.figure(figsize = (10, 4))
axis = curves_figure.add_subplot(1, 2, 1)
epochs_plotted = plot_curve(axis, parameters, train_column = "train_accuracy_history", valid_column = "valid_accuracy_history")
pyplot.grid()
pyplot.legend()
pyplot.xlabel("epoch")
pyplot.ylabel("accuracy")
pyplot.ylim(50., 115.)
pyplot.xlim(0, epochs_plotted)#max(new_model_epochs, old_model_epochs))
axis = curves_figure.add_subplot(1, 2, 2)
epochs_plotted = plot_curve(axis, parameters, train_column = "train_loss_history", valid_column = "valid_loss_history")
pyplot.grid()
pyplot.legend()
pyplot.xlabel("epoch")
pyplot.ylabel("loss")
pyplot.ylim(0.0001, 10.)
pyplot.xlim(0, epochs_plotted)
pyplot.yscale("log")
The actual method to perform model training. Calling this function would trigger the training process.
In [14]:
def train_model(params, X_train, y_train, X_valid, y_valid, X_test, y_test, logger_config):
"""
Performs model training based on provided training dataset
according to provided parameters, and then evaluates trained
model with testing dataset.
Part of the training dataset may be used for validation during
training if specified in model parameters.
Parameters
----------
params : Parameters
Structure (`namedtuple`) containing model parameters.
X_train :
Training dataset.
y_train :
Training dataset labels.
X_valid :
Validation dataset.
y_valid :
Validation dataset labels.
X_test :
Testing dataset.
y_test :
Testing dataset labels.
"""
# Initialisation routines: generate variable scope, create logger, note start time.
paths = Paths(params)
log = ModelCloudLog(
os.path.join(paths.root_path, "logs"),
dropbox_token=logger_config["dropbox_token"],
telegram_token=logger_config["telegram_token"],
telegram_chat_id=logger_config["telegram_chat_id"]
)
start = time.time()
model_variable_scope = paths.var_scope
log.log_parameters(params, y_train.shape[0], y_valid.shape[0], y_test.shape[0])
# Build the graph
graph = tf.Graph()
with graph.as_default():
# Input data. For the training data, we use a placeholder that will be fed at run time with a training minibatch.
tf_x_batch = tf.placeholder(tf.float32, shape = (None, params.image_size[0], params.image_size[1], 1))
tf_y_batch = tf.placeholder(tf.float32, shape = (None, params.num_classes))
is_training = tf.placeholder(tf.bool)
current_epoch = tf.Variable(0, trainable=False) # count the number of epochs
# Model parameters.
if params.learning_rate_decay:
learning_rate = tf.train.exponential_decay(params.learning_rate, current_epoch, decay_steps = params.max_epochs, decay_rate = 0.01)
else:
learning_rate = params.learning_rate
# Training computation.
with tf.variable_scope(model_variable_scope):
logits = model_pass(tf_x_batch, params, is_training)
if params.l2_reg_enabled:
with tf.variable_scope('fc4', reuse = True):
l2_loss = tf.nn.l2_loss(tf.get_variable('weights'))
else:
l2_loss = 0
predictions = tf.nn.softmax(logits)
softmax_cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits, tf_y_batch)
loss = tf.reduce_mean(softmax_cross_entropy) + params.l2_lambda * l2_loss
# Optimizer.
optimizer = tf.train.AdamOptimizer(
learning_rate = learning_rate
).minimize(loss)
with tf.Session(graph = graph) as session:
session.run(tf.global_variables_initializer())
# A routine for evaluating current model parameters
def get_accuracy_and_loss_in_batches(X, y):
p = []
sce = []
batch_iterator = BatchIterator(batch_size = 128)
for x_batch, y_batch in batch_iterator(X, y):
[p_batch, sce_batch] = session.run([predictions, softmax_cross_entropy], feed_dict = {
tf_x_batch : x_batch,
tf_y_batch : y_batch,
is_training : False
}
)
p.extend(p_batch)
sce.extend(sce_batch)
p = np.array(p)
sce = np.array(sce)
accuracy = 100.0 * np.sum(np.argmax(p, 1) == np.argmax(y, 1)) / p.shape[0]
loss = np.mean(sce)
return (accuracy, loss)
# If we chose to keep training previously trained model, restore session.
if params.resume_training:
try:
tf.train.Saver().restore(session, paths.model_path)
except Exception as e:
log("Failed restoring previously trained model: file does not exist.")
pass
saver = tf.train.Saver()
early_stopping = EarlyStopping(tf.train.Saver(), session, patience = params.early_stopping_patience, minimize = True)
train_loss_history = np.empty([0], dtype = np.float32)
train_accuracy_history = np.empty([0], dtype = np.float32)
valid_loss_history = np.empty([0], dtype = np.float32)
valid_accuracy_history = np.empty([0], dtype = np.float32)
if params.max_epochs > 0:
log("================= TRAINING ==================")
else:
log("================== TESTING ==================")
log(" Timestamp: " + get_time_hhmmss())
log.sync()
for epoch in range(params.max_epochs):
current_epoch = epoch
# Train on whole randomised dataset in batches
batch_iterator = BatchIterator(batch_size = params.batch_size, shuffle = True)
for x_batch, y_batch in batch_iterator(X_train, y_train):
session.run([optimizer], feed_dict = {
tf_x_batch : x_batch,
tf_y_batch : y_batch,
is_training : True
}
)
# If another significant epoch ended, we log our losses.
if (epoch % params.log_epoch == 0):
# Get validation data predictions and log validation loss:
valid_accuracy, valid_loss = get_accuracy_and_loss_in_batches(X_valid, y_valid)
# Get training data predictions and log training loss:
train_accuracy, train_loss = get_accuracy_and_loss_in_batches(X_train, y_train)
if (epoch % params.print_epoch == 0):
log("-------------- EPOCH %4d/%d --------------" % (epoch, params.max_epochs))
log(" Train loss: %.8f, accuracy: %.2f%%" % (train_loss, train_accuracy))
log("Validation loss: %.8f, accuracy: %.2f%%" % (valid_loss, valid_accuracy))
log(" Best loss: %.8f at epoch %d" % (early_stopping.best_monitored_value, early_stopping.best_monitored_epoch))
log(" Elapsed time: " + get_time_hhmmss(start))
log(" Timestamp: " + get_time_hhmmss())
log.sync()
else:
valid_loss = 0.
valid_accuracy = 0.
train_loss = 0.
train_accuracy = 0.
valid_loss_history = np.append(valid_loss_history, [valid_loss])
valid_accuracy_history = np.append(valid_accuracy_history, [valid_accuracy])
train_loss_history = np.append(train_loss_history, [train_loss])
train_accuracy_history = np.append(train_accuracy_history, [train_accuracy])
if params.early_stopping_enabled:
# Get validation data predictions and log validation loss:
if valid_loss == 0:
_, valid_loss = get_accuracy_and_loss_in_batches(X_valid, y_valid)
if early_stopping(valid_loss, epoch):
log("Early stopping.\nBest monitored loss was {:.8f} at epoch {}.".format(
early_stopping.best_monitored_value, early_stopping.best_monitored_epoch
))
break
# Evaluate on test dataset.
test_accuracy, test_loss = get_accuracy_and_loss_in_batches(X_test, y_test)
valid_accuracy, valid_loss = get_accuracy_and_loss_in_batches(X_valid, y_valid)
log("=============================================")
log(" Valid loss: %.8f, accuracy = %.2f%%)" % (valid_loss, valid_accuracy))
log(" Test loss: %.8f, accuracy = %.2f%%)" % (test_loss, test_accuracy))
log(" Total time: " + get_time_hhmmss(start))
log(" Timestamp: " + get_time_hhmmss())
# Save model weights for future use.
saved_model_path = saver.save(session, paths.model_path)
log("Model file: " + saved_model_path)
np.savez(paths.train_history_path, train_loss_history = train_loss_history, train_accuracy_history = train_accuracy_history, valid_loss_history = valid_loss_history, valid_accuracy_history = valid_accuracy_history)
log("Train history file: " + paths.train_history_path)
log.sync(notify=True, message="Finished training with *%.2f%%* accuracy on the testing set (loss = *%.6f*)." % (test_accuracy, test_loss))
plot_learning_curves(params)
log.add_plot(notify=True, caption="Learning curves")
pyplot.show()
Here we preprocess the whole dataset first. We first flip some of the images which classes allow that, and then prepare two versions of the training dataset: extended and balanced. Extended one has the same distribution of examples across classes, but has 20x more augmented data. Balanced contains examples equally balanced across classes. In both cases we don't simply copy the images — we apply augmentation, so that every time we duplicate the data it's slightly different.
Prepare extended and balanced training datasets by augmenting original data:
In [ ]:
import random
import pickle
from sklearn.cross_validation import train_test_split
train_dataset_file = "traffic-signs-data/train.p"
test_dataset_file = "traffic-signs-data/test.p"
train_extended_dataset_file = "traffic-signs-data/train_extended.p"
train_balanced_dataset_file = "traffic-signs-data/train_balanced.p"
X_train, y_train = load_pickled_data(train_dataset_file, ['features', 'labels'])
print("Number of training examples in initial dataset =", X_train.shape[0])
_, class_counts = np.unique(y_train, return_counts = True)
X_train, y_train = flip_extend(X_train, y_train)
print("Number of training examples after horizontal flipping =", X_train.shape[0])
# Prepare a dataset with balanced classes
X_train_balanced, y_train_balanced = extend_balancing_classes(X_train, y_train, aug_intensity = 0.75, counts = np.full(43, 20000, dtype = int))
print("Number of training examples after augmenting and balancing training data =", X_train_balanced.shape[0])
pickle.dump({
"features" : X_train_balanced,
"labels" : y_train_balanced
}, open(train_balanced_dataset_file, "wb" ) )
print("Balanced dataset saved in", train_balanced_dataset_file)
# Prepare a dataset with extended classes
X_train_extended, y_train_extended = extend_balancing_classes(X_train, y_train, aug_intensity = 0.75, counts = class_counts * 20)
print("Number of training examples after augmenting and extending training data =", X_train_extended.shape[0])
pickle.dump({
"features" : X_train_extended,
"labels" : y_train_extended
}, open(train_extended_dataset_file, "wb" ) )
print("Extended dataset saved in", train_extended_dataset_file)
Preprocess all datasets:
In [ ]:
import pickle
train_extended_dataset_file = "traffic-signs-data/train_extended.p"
train_balanced_dataset_file = "traffic-signs-data/train_balanced.p"
train_extended_preprocessed_dataset_file = "traffic-signs-data/train_extended_preprocessed.p"
train_balanced_preprocessed_dataset_file = "traffic-signs-data/train_balanced_preprocessed.p"
test_dataset_file = "traffic-signs-data/test.p"
test_preprocessed_dataset_file = "traffic-signs-data/test_preprocessed.p"
X_train, y_train = load_and_process_data(train_balanced_dataset_file)
pickle.dump({
"features" : X_train,
"labels" : y_train
}, open(train_balanced_preprocessed_dataset_file, "wb" ) )
print("Preprocessed balanced training dataset saved in", train_balanced_preprocessed_dataset_file)
X_train, y_train = load_and_process_data(train_extended_dataset_file)
pickle.dump({
"features" : X_train,
"labels" : y_train
}, open(train_extended_preprocessed_dataset_file, "wb" ) )
print("Preprocessed extended training dataset saved in", train_extended_preprocessed_dataset_file)
X_test, y_test = load_and_process_data(test_dataset_file)
pickle.dump({
"features" : X_test,
"labels" : y_test
}, open(test_preprocessed_dataset_file, "wb" ) )
print("Preprocessed extended testing dataset saved in", test_preprocessed_dataset_file)
Specify model hyperparameters and perform training.
In [328]:
import pickle
import json
from sklearn.model_selection import train_test_split
train_extended_preprocessed_dataset_file = "traffic-signs-data/train_extended_preprocessed.p"
train_balanced_preprocessed_dataset_file = "traffic-signs-data/train_balanced_preprocessed.p"
test_preprocessed_dataset_file = "traffic-signs-data/test_preprocessed.p"
with open(os.path.join(os.getcwd(), 'config.json')) as data_file:
logger_config = json.load(data_file)
parameters = Parameters(
# Data parameters
num_classes = 43,
image_size = (32, 32),
# Training parameters
batch_size = 256,
max_epochs = 1001,
log_epoch = 1,
print_epoch = 1,
# Optimisations
learning_rate_decay = False,
learning_rate = 0.0001,
l2_reg_enabled = True,
l2_lambda = 0.0001,
early_stopping_enabled = True,
early_stopping_patience = 100,
resume_training = True,
# Layers architecture
conv1_k = 5, conv1_d = 32, conv1_p = 0.9,
conv2_k = 5, conv2_d = 64, conv2_p = 0.8,
conv3_k = 5, conv3_d = 128, conv3_p = 0.7,
fc4_size = 1024, fc4_p = 0.5
)
X_train, y_train = load_pickled_data(train_balanced_preprocessed_dataset_file, columns = ['features', 'labels'])
X_train, X_valid, y_train, y_valid = train_test_split(X_train, y_train, test_size = 0.25)
X_test, y_test = load_pickled_data(test_preprocessed_dataset_file, columns = ['features', 'labels'])
train_model(parameters, X_train, y_train, X_valid, y_valid, X_test, y_test, logger_config)
The accuracy of 99.33% means that there were 85 images in the test set that this model failed to classify correctly — let's take a look at those 85 images.
In [15]:
def get_predictions(params, X):
"""
Evaluates `X` on a model defined by `params`.
Parameters
----------
params : Parameters
Structure (`namedtuple`) containing model parameters.
X :
Testing dataset.
"""
# Initialisation routines: generate variable scope, create logger, note start time.
paths = Paths(params)
# Build the graph
graph = tf.Graph()
with graph.as_default():
# Input data. For the training data, we use a placeholder that will be fed at run time with a training minibatch.
tf_x = tf.placeholder(tf.float32, shape = (None, params.image_size[0], params.image_size[1], 1))
is_training = tf.constant(False)
with tf.variable_scope(paths.var_scope):
predictions = tf.nn.softmax(model_pass(tf_x, params, is_training))
top_5_predictions = tf.nn.top_k(predictions, 5)
with tf.Session(graph = graph) as session:
session.run(tf.global_variables_initializer())
tf.train.Saver().restore(session, paths.model_path)
[p] = session.run([top_5_predictions], feed_dict = {
tf_x : X
}
)
return np.array(p)
In [34]:
X_test, y_test = load_pickled_data(test_preprocessed_dataset_file, columns = ['features', 'labels'])
X_original, _ = load_pickled_data(test_dataset_file, columns = ['features', 'labels'])
predictions = get_predictions(parameters, X_test)
predictions = predictions[1][:, np.argmax(predictions[0], 1)][:, 0].astype(int)
labels = np.argmax(y_test, 1)
print("Original:")
incorrectly_predicted = X_original[predictions != labels]
fig = pyplot.figure(figsize=(6, 6))
fig.subplots_adjust(left=0, right=1, bottom=0, top=1, hspace=0.05, wspace=0.05)
for i in range(incorrectly_predicted.shape[0]):
ax = fig.add_subplot(10, 10, i + 1, xticks=[], yticks=[])
ax.imshow(incorrectly_predicted[i])
pyplot.show()
print("Preprocessed:")
incorrectly_predicted = X_test[predictions != labels]
fig = pyplot.figure(figsize=(6, 6))
fig.subplots_adjust(left=0, right=1, bottom=0, top=1, hspace=0.05, wspace=0.05)
for i in range(incorrectly_predicted.shape[0]):
ax = fig.add_subplot(10, 10, i + 1, xticks=[], yticks=[])
ax.imshow(incorrectly_predicted[i].reshape(32, 32), cmap='gray')
pyplot.show()
Looks like most of them had artefacts like sunlight or obstructing objects, or didn't have enough data for in the training set. This is one of the issues of training the model on unbalanced data: probably those 100 and 60 signs were incorrectly classified as 80, as we had significantly more data for those after flipping.
Step 3: Test a Model on New Images¶
Try out trained model on manually captured images (in Google Maps).
In [330]:
import matplotlib.gridspec as gridspec
from skimage import io
import os
def plot_image_statistics(predictions, index):
# Prepare original and preprocessed images
original = io.imread(os.getcwd() + '/traffic-signs-data/custom/' + "example_{0:0>5}".format(index + 1) + '.png')
preprocessed = X_custom[index].reshape(32, 32)
# Prepare the grid
pyplot.figure(figsize = (6, 2))
gridspec.GridSpec(2, 2)
# Plot original image
pyplot.subplot2grid((2, 2), (0, 0), colspan=1, rowspan=1)
pyplot.imshow(original)
pyplot.axis('off')
# Plot preprocessed image
pyplot.subplot2grid((2, 2), (1, 0), colspan=1, rowspan=1)
pyplot.imshow(preprocessed, cmap='gray')
pyplot.axis('off')
# Plot predictions
pyplot.subplot2grid((2, 2), (0, 1), colspan=1, rowspan=2)
pyplot.barh(np.arange(5)+.5, predictions[0][index], align='center')
pyplot.yticks(np.arange(5)+.5, signnames[predictions[1][index].astype(int)])
pyplot.tick_params(axis='both', which='both', labelleft='off', labelright='on', labeltop='off', labelbottom='off')
pyplot.show()
# Load images from .png files to `X_custom` NumPy array
X_custom = np.empty([0, 32, 32, 3], dtype = np.int32)
for i in range(38):
image = io.imread(os.getcwd() + '/traffic-signs-data/custom/' + "example_{0:0>5}".format(i + 1) + '.png')
X_custom = np.append(X_custom, [image[:, :, :3]], axis = 0)
# Preprocess and get predictions based on the current model
X_custom, _ = preprocess_dataset(X_custom)
predictions = get_predictions(parameters, X_custom)
# For every example plot original image, preprocessed image, and model predictions.
for i in range(38):
plot_image_statistics(predictions, i)
Question 1¶
Describe the techniques used to preprocess the data.
Answer:
- Images were converted to grayscale by taking Y component of their YCbCr represenation.
- Images had localised histogram equalisation applied. It seemed to extract features better than plain histogram equalisation.
- Pixel values were scaled to
[0, 1].
Question 2¶
Describe how you set up the training, validation and testing data for your model. If you generated additional data, why?
Answer:
I decided to use a 25% slice of training data for validation. Normally I would split entire dataset as 70% training, 15% validation, 15% testing, but since here test dataset was supplied separately I assumed it's better to use all of it solely for testing.
The dataset looked quite small and some classes were significantly less represented than others, so I reckon the model wouldn't generalise well without additional examples. I decided to use data augmentation by applying randomised transformations. I find Lasagne's BatchIterator class very convenient, as one can simply override the transform() function and provide their data augmentation pipeline there.
So that's what I did, my pipeline includes:
- Flip. This is the first step before any further augmentation, as it lets us extend some of the datasets significantly. On this step we flip images horizontally, vertically, or horizontally+vertically (e.g. rotate) if their class would remain the same, OR flips and updates the label, if flipped image represents some other class (e.g. flipped Turn left sign would become Turn right).
- Rotation. Rotates the image at a random angle in the range of
[-15, 15]degrees at most. - Projection transform. Warps the image by projecting each of the image corners to a random point within a ~10% of image size distance from the original point at most.
I also experimented with other adjustments, but they didn't seem to improve model performance, so removed them from the final model. Those were:
- Gamma adjustment. Adjusts image brightness/contrast by applying gamma adjustment in the range of
[0.5, 1.5]. - Blur. Applies gaussian blur with sigma in the range of
[0, 0.8]. - Noise. Adds random noise to the image. I've switched it off for now, but might experiment with it later.
I chose these transformations as they seemed to be close to what we could see on the photos. Image could be blurred and/or noisy due to low resolution source photo, it can be either too bright or too dark depending on the outside conditions, and it can look rotated and/or projection warped on the photo depending on the shooting angle.
Question 3¶
What does your final architecture look like? (Type of model, layers, sizes, connectivity, etc.) For reference on how to build a deep neural network using TensorFlow, see Deep Neural Network in TensorFlow from the classroom.
Answer:
I used a fairly simple model of 3 convolutional layers and 1 fully connected layer (with every convolutional and fully connected layer being followed by a ReLU non-linearity):
As per Pierre Sermanet paper, I have provided output of every convolutional layer to the classifier, applying additional pooling to each of the convolutinal outputs, so that they are proportionally downsampled.
Question 4¶
How did you train your model? (Type of optimizer, batch size, epochs, hyperparameters, etc.)
Answer:
I used AdamOptimizer as it seemed to perform quite well in similar tasks. I experimented with various hyperparameters, including batch size, epochs and various regularisation techniques, and ended up using batches of 256 examples, applying L2, dropout and early stopping for regularisation. I'm only applying L2 to fully connected layers and am trying to minimise dropout on convolutional layers:
--------------- ARCHITECTURE ----------------
Type Size Dropout (keep p)
Layer 1 5x5 Conv 32 0.9
Layer 2 5x5 Conv 64 0.8
Layer 3 5x5 Conv 128 0.7
Layer 4 FC 1024 0.5
---------------- PARAMETERS -----------------
L2 Regularization: Enabled (lambda = 0.0001)
Early stopping: Enabled (patience = 100)Training itself was carried out in two stages:
Pretraining stage¶
Train on a extended training set (having 20x more data) with the following parameters:
- Learning rate = 0.001
- Augmentation intensity = 0.75
This stage took ~3.5 hours on a GTX1080.
Additional fine-tuning stages¶
Train on a balanced training set (having 20.000 examples for each class) with a decreased learning rate:
- Learning rate = 0.0001
- Augmentation intensity = 0.75
After these two steps one can easily get ~99% accuracy on the testing set. However, I then repeated this fine-tuning training a couple of more times by re-generating balanced dataset and gradually reducing augmentation intensity, which eventually got me the highest of 99.33% accuracy on the testing set.
Question 5¶
What approach did you take in coming up with a solution to this problem?
Answer:
My solution was inspired by Daniel Nouri's tutorial and Pierre Sermanet paper provided in the course materials.
Question 6¶
Choose five candidate images of traffic signs and provide them in the report. Are there any particular qualities of the image(s) that might make classification difficult? It would be helpful to plot the images in the notebook.
Answer:
I have collected images by "walking" in my neighborhood with Google Maps, as I assumed this would be most realistic examples (taken from an actual moving car).
I've collected data in different cities, and the ones collected in London were pretty accurate in predictions, as they look similar to German signs. Accuracy of classifying examples collected in other countries was significantly lower, since some signs look fairly different. So the model does not generalise very well on new types pf signs.
A good example would be a couple of signs incorrectly classified as Priority road:
In [331]:
for i in [2, 14, 19]:
plot_image_statistics(predictions, i)
Surprisingly, the model was quite sure about some of these predictions. My best guesses are:
- These last two signs look fairly different from the German ones.
- First training stage was carried out on a extended dataset and not balanced one, the model became biased towards predicting overrepresented classes. Original training dataset contained twice as many examples for Priority road class than for No entry, so it naturally became biased towards predicting the class it has seen many examples of during training.
Question 7¶
Is your model able to perform equally well on captured pictures when compared to testing on the dataset?
Answer:
It's pretty close (and confident about its predictions), when signs are similar to German ones, and also when there are no distortions.
If, however, signs look slightly different and/or are damaged, the model does not predict very well and can be biased towards signs that were overrepresented in the training dataset.
Unfortunately this model kind of overfits the supplied test data (although I didn't use it while training, of course), as my main goal was to get highest score on the supplied testing set. If I were to implement the model that should perform best on arbitrary captured dataset, I would only train it on class-balanced data, where all classes were equally represented. However, since the goal was to get highest test set accuracy, it may be overfitting it a bit.
Question 8¶
Use the model's softmax probabilities to visualize the certainty of its predictions, tf.nn.top_k could prove helpful here. Which predictions is the model certain of? Uncertain? If the model was incorrect in its initial prediction, does the correct prediction appear in the top k? (k should be 5 at most)
Answer:
It is fairly certain about predictions of signs that look similar to German ones. It does not generalise well on signs from other countries, although in most cases the correct sign is in top 5.
Question 9¶
If necessary, provide documentation for how an interface was built for your model to load and classify newly-acquired images.
Answer:
I didn't create any additional interface and instead put the images into a folder with preformatted file names, manually loading them and running through a TensorFlow graph with restored trained weights.